4X4 matrix multiplication using MMX instructions

void mult4x4(short m1[4][4], short m2[4][4], short res[4][4])
{
	__asm
	{
		push		esi
		push		edi

		mov		eax,	m1
		mov		esi,	m2
		mov		edi,	res

		; cache mem. access into free mmx reg.
		movq		mm5,	[esi + 24]

		mov		edx,	8
		mov		ecx,	4
__repeat:
		movq		mm7,	[eax]
		movq		mm0,	[esi]
		movq		mm2,	[esi + edx]
		movq		mm1,	mm0

		movq		mm6,	mm7
		punpckldq	mm7,	mm6

		punpcklwd	mm0,	mm2
		punpckhwd	mm1,	mm2

		pmaddwd		mm0,	mm7
		pmaddwd		mm1,	mm7

		movq		mm3,	mm0
		movq		mm4,	mm1

		movq		mm7,	mm6
		punpckhdq	mm7,	mm6

		movq		mm0,	[esi + edx * 2]
		movq		mm2,	mm5  ; read row from reg. cache
		movq		mm1,	mm0

		punpcklwd	mm0,	mm2
		punpckhwd	mm1,	mm2

		pmaddwd		mm0,	mm7
		pmaddwd		mm1,	mm7

		paddd		mm3,	mm0
		paddd		mm4,	mm1

		packssdw	mm3,	mm3
		packssdw	mm4,	mm4
		punpckldq	mm3,	mm4
		movq		[edi],	mm3

		add		edi,	edx
		add		eax,	edx

		loop		__repeat

		emms

		pop		edi
		pop		esi
	}
}

int main(void)
{
    __declspec(align(16)) short m1[4][4] = { 1,1,1,1,2,2,2,2,3,3,3,3,4,4,4,4 };
    __declspec(align(16)) short m2[4][4] = { 9,9,9,9,8,8,8,8,7,7,7,7,6,6,6,6 };
    __declspec(align(16)) short result[4][4] = {0};

    mult4x4(m1, m2, result);

    for(int r = 0; r < 4; r++)
    {
        for(int c = 0; c < 4; c++)
            cout << "\t" << result[r][c];
        cout << endl;
    }

    cout << endl;

    return 0;
}